Agent Governance for Government and Defense Teams

The first question in a government or defense evaluation is placement: can the software run where we already run? Modulo is self-hosted, so the answer starts with your infrastructure and your network boundary.

Placement

Modulo runs on Docker Compose, on a Linux host through a native installer that bundles PostgreSQL 16, Redis 8, and a Python runtime, or on a VM or bare-metal host you manage. There is no hosted edition and no cloud dependency. A deployment can sit on a subnet with no route to the public internet.

Egress you authorize

With default settings and no connectors configured, Modulo makes no external network calls. There is no telemetry endpoint and no phone-home for license validation: every outbound connection is one you configure, whether that is a model backend, a connector, a notification webhook, or an identity provider.

Model and agent placement

Model calls follow the backends you attach. For inference that stays on your network, a local OpenAI-compatible server is a supported backend. For the tool-using loop, the Bundled Runner executes sandbox_agent nodes as hardened containers on your own Docker engine, behind a filtered Docker socket proxy and on a dedicated workspace network. A profile set to network_policy: none gives a workspace loopback-only networking.

The record your assessors read

Every state-changing action appends an immutable audit event, and each event carries the SHA-256 hash of the one before it, so the chain is checkable end to end. Events are append-only at the application layer and at the database layer. The chain lives in your Postgres database, where your team controls retention, backup, and export, and a verification endpoint recomputes it on demand. Each run also records the prompt, the output, the schema result, and the gate decisions.

Access control

Single sign-on over OIDC or SAML, just-in-time account creation, identity-provider group mapping to teams, and four organization roles (admin, operator, runner, viewer) are the access-control surface, all self-hosted. Single sign-on, team roles, and the audit viewer and export are part of the paid Team tier; the Community tier runs the pipelines and records the audit chain without them. Tier detail is on pricing.

Accreditation is an assessment of the deployment you assemble

Frameworks like FedRAMP and the IL levels assess a deployment: where the data sits, where the network boundary is, who holds the keys, and who can operate the system. On self-hosted software those decisions are your side of the line. How that split works across sectors: Agent governance for regulated engineering teams.